Loop detection

Goa has several independent safeguards for runaway model behaviour. They do not share a single counter, so changing one setting does not change the others. Values below are configured under execution in YAML (or through /config where noted). A value of 0 means the documented default unless stated otherwise.

Tool-call repeat detector

This detector watches identical tool calls (same tool and arguments) and reports a loop when they repeat consecutively.

Setting	Default	Purpose
`loop_warning`	10	Same tool-call streak that produces a warning.
`loop_interrupt`	15	Same tool-call streak that interrupts execution. Must exceed `loop_warning`.
`max_tool_repeat_total`	off	Maximum identical calls in a turn, regardless of spacing.
`max_tool_repeat_consecutive`	2	Maximum consecutive identical calls in a turn.
`max_tool_calls`	200	Maximum occurrences of one identical call in the rolling window. This is not a total tool-call allowance.
`tool_call_limit_reset_window`	10	Number of recent calls considered by the rolling-window guard.
`disable_tool_budget`	false	Disables the agent tool-call budget guards. It does not disable the core loop detector.

The interactive path is /config → Loop detection → Thresholds. The same values can be set with /config:set:execution.<key>:<value>.

Consecutive tool-call rounds

max_consecutive_tool_rounds (default 15) limits consecutive LLM rounds that end in tool calls. It catches loops where every call is different, so duplicate-call guards never fire. When reached, Goa injects an internal recovery instruction asking the model to answer with the information already gathered. This is the source of the “per-turn tool-call round limit” message.

It is available in /config → Loop detection → Thresholds. Set it to 0 to disable it. This is separate from max_tool_calls: the embedded default is 200, and changing that duplicate-call limit does not raise this round limit.

Thinking/reasoning detector

The thinking detector tracks repeated significant reasoning lines and low-diversity reasoning windows within one turn.

Setting	Purpose
`disable_thinking_loop_detection`	Persistently disable thinking-loop detection.
`disable_thinking_stall_detection`	Persistently disable the reasoning-only stall watchdog.
`thinking_stall_warn_seconds`	Delay before a reasoning-only stall warning.
`thinking_stall_stop_seconds`	Delay before stopping a reasoning-only stall.

Thinking-loop thresholds use the runtime loop-detector defaults and are not currently separate YAML fields. Session-only overrides are available with /config:temp:think_loop_detection:on|off and /config:temp:thinking_stall_detection:on|off.

Streaming text detector

This detector catches repeated text in the assistant's streamed output, even when no tool is called.

Setting	Default	Purpose
`stream_loop_max_repeats`	5	Repeated blocks required before stopping the stream.
`stream_loop_min_period`	50	Minimum repeated unit length in characters. Values below 8 are rejected.
`stream_loop_max_strikes`	3	Loop detections before stopping the turn.
`stream_loop_reset_after`	10	Clean messages/tool calls before resetting strikes.
`disable_stream_loop_detection`	false	Persistently disable this detector.

Session-only override: /config:temp:stream_loop_detection:on|off.

Loop auto-resume

By default, when a loop detector (thinking or tool) interrupts the turn, the agent stops and waits for the user. Loop auto-resume keeps the stop but automatically sends a follow-up message so the agent resumes instead of staying idle.

Setting	Default	Description
`loop_auto_resume`	false	Send a follow-up message after a loop-detector stop so the agent resumes.
`loop_auto_resume_message`	`loop detected and you were stopped - resume now`	The message injected as a user turn on auto-resume.
`loop_auto_resume_max`	3	Maximum consecutive loop-triggered auto-resumes before giving up. Each genuine user turn resets the counter, so a stuck agent cannot resume forever.

The interactive path is /config → Loop detection → Loop auto-resume, which exposes the on/off toggle and the resume-message editor. The same values can be set with /config:set:execution.loop_auto_resume:true and /config:set:execution.loop_auto_resume_message:<text>.

If you typed steering input while the looped turn was still running, that steering is delivered instead of the automated resume — your message takes precedence.

Cross-turn repeated-response guardrail

This detector watches the assistant's last message of each turn and stops the session when the same response repeats across consecutive turns. Goal continuation turns deliberately keep its counters, so a driver loop latches; a genuine user turn resets it.

Setting	Default	Purpose
`runaway_loop_max_repeats`	2	Repeats after the first occurrence before the session stops.

Progress overrides repetition: a turn whose tool calls produced non-error results never accumulates a strike, whatever its text. A truly empty response ((empty response): no text, no reasoning, no tool calls) is judged against the whole turn — if that turn executed successful tool calls, the emptiness is a provider stall (an empty stop right after tool results), not the model repeating itself, and the turn reports a silent stop instead (see the goal pause reason “stopped without producing a reply”). Only turns that produced nothing at all still latch.

Configuration examples

execution:
  loop_warning: 10
  loop_interrupt: 15
  max_tool_repeat_total: 15
  max_tool_repeat_consecutive: 2
  max_tool_calls: 200
  tool_call_limit_reset_window: 10
  max_consecutive_tool_rounds: 15
  stream_loop_max_repeats: 5
  runaway_loop_max_repeats: 2
  loop_auto_resume: false
  loop_auto_resume_message: "loop detected and you were stopped - resume now"
  loop_auto_resume_max: 3

If a warning appears sooner than expected, identify which detector produced it. In particular, max_tool_calls controls duplicate calls in a rolling window; it does not control consecutive tool-call rounds.